Automated QT Interval Analysis on Diagnostic ... - PhysioNet

<strong>Automated</strong> <strong>QT</strong> <strong>Interval</strong> <strong>Analysis</strong>
on Diagnostic Electrocardiograms
DA Tironi 1 , R Sassi 1 , LT Mainardi 2
1 Dipartimento di Tecnologie dell'Informazione, Università di Milano, Crema, Italy
2 Dipartimento di Bioingegneria, Politecnico di Milano, Milano, Italy
Abstract
In this study, we suggest a method for automatic
measurement of the <strong>QT</strong> interval. The method derives from
the standard technique of “selective beat averaging”. To
overcome those situations in which the templates are
noisy as they arise from averaging only a small number
of beats, two separate patterns were used: one for the
QRS complex and the other for the T wave. The templates
were carefully matched with the beats by crosscorrelation,
using an iterative refinement process.
The method was employed to automatically measure
<strong>QT</strong> segments on the PTB Diagnostic ECG Database used
for the 2006 Computers in Cardiology Challenge. A score
in division 3 of 81.07 ms was obtained pointing out that
the method needs further tuning and refinements.
Excluding those beats for which the estimated <strong>QT</strong> values
are outside the main physiological range (290-500 ms) a
more promising tentative score of 41.62 ms might be
obtained.
1. Introduction
The <strong>QT</strong> interval, e.g. the time elapsed between the Q
onset and the end of the T wave, as measured on a
standard diagnostic electrocardiogram, is typically
employed to derive an indirect measure of the cardiac
repolarization period. A prolonged cardiac repolarization
was shown to be a prominent predictor of development of
fatal cardiac arrhythmias [1][2] as it typically lead to fatal
arrhythmias, like torsade de pointes. Among others,
factors inducing prolonged <strong>QT</strong> intervals are: myocarditis
[3][4], certain diseases in which electrolyte imbalance
play a role, like hypocalcemia [4][5] and certain drugs
(e.g. Quinidine, Sotalol) [6]. As pointed out by [6] “in the
past decade, the single most common cause of the
withdrawal or restriction of the use of drugs that have
already been marketed has been the prolongation of the
<strong>QT</strong> interval.” It is no surprise then that a precise measure
of the <strong>QT</strong> interval is fundamental for the diagnostic
procedure of several diseases and for pharmaceutical
trials.
The <strong>QT</strong> interval quantification is complicated by the
fact that the repolarization period depends on the heart
rate and it shortens when the frequency increases. Starting
from the work of Bazett [7], several formulas have been
proposed to model the <strong>QT</strong>/RR relationship, but as
recently suggested such relation is highly individual [8].
A second possible approach is that of selecting a
representative beat, a sort of median value, but the criteria
underlying such choice are often subjective to the use of
the measure envisioned by the physician.
The option to automatically measure the <strong>QT</strong> interval
on long-term computerized ECG is today available on
most commercial systems. Even so, manual
measurements of <strong>QT</strong> intervals on ECG tracings are de
facto the only generally accepted procedure for a
quantitative evaluation. For example, the recently issued
guidelines (ICH E14, [9]) for clinical evaluation of nonantiarrhythmic
drugs, while calling for “thorough
<strong>QT</strong>/<strong>QT</strong>c studies”, endorse only manual <strong>QT</strong> interval
measurements. This is possibly due to the lack of
precision that automated methods shown in the past.
Manual measurements bear though well-known
limitations [10]: they are affected by the paper speed used
to plot the electrocardiogram and by electrocardiogram
gain (T amplitude); they vary with different readings and
readers (on average up to 20 ms). Lastly, they need to be
performed by a trained physician and they become
unfeasible on long recordings where only a few indicative
beats are thus taken into account.
Surely enough the availability of automatic procedures
for the measurements of the <strong>QT</strong> interval without the
intervention of the physician and with a precision similar
to manual readings would foster new physiological
researches on the ventricular repolarization period on
long term Holter recordings and might also permit more
exhaustive protocols during pharmacological and clinical
trials. In this work we suggest such an algorithm which
building over a standard technique, called “RR bin” or
“selective beat averaging” [11] might be able to provide a
reliable estimate of the <strong>QT</strong> interval. The method was
tested on the PTB Diagnostic ECG Database which forms
the testset of the 2006 PhysioNet/Computers in

Cardiology Challenge [12]. The database consists of 549
recordings of a couple of minutes each, collected from
294 subjects. For each record the 12 conventional leads
and the 3 XYZ leads had been made available.
2. Methods
2.1. Preprocessing
The dataset did not contain annotations. Thus, each
record was firstly processed for QRS detection using the
freely available software ECGPUWAVE [13]. The
routine also provided an estimate of the waveforms’
boundaries which were used as starting point of the
method described below. The choice was made just for
convenience (the code was faster) and it could have been
completely avoided using a fixed <strong>QT</strong> initial guess.
2.2. <strong>QT</strong> measurement
The ECG signal was first high-pass filtered to remove
major baseline wander (5 th order Butterworth filter, cut
off frequency 0.5 Hz).
Then local average patterns were built subject to the
condition that the RR interval between successive beats
did non change more than a specified threshold. In fact,
as previously outlined, the <strong>QT</strong> interval changes with the
heart rate; more importantly the <strong>QT</strong> length needs to adapt
to the new pace when it changes and the process takes
time. Building average patterns using only beats which
are about at the same distance, it ensures that the <strong>QT</strong>
measures might be taken in a portion of the ECG where
<strong>QT</strong> adaptation already occurred (and measures are thus
reproducible). Specifically the local patterns were built
through successive refinement in an iterative process.
First, within each ECG recording the groups of beat for
which |RR i+1 -RR i |

of beats belonging to each group). The iterative process
was ended when the number of adjustments in a cycle
was small or the number of cycles exceeded a threshold.
To speed up the refinement process, instead of using a
first T end guess located 500 ms from the Q start, we also
employed the <strong>QT</strong> estimate provided by ECGPWAVE as
specified in section 2.1. A couple of local average
patterns obtained from a group of four beats are shown in
figure 1.
Once adjusted the position of the two conventional
points with respect to the average patterns and thus
obtained patterns which reflected the morphology of each
group more carefully, the Q onset and T offset were
measured on the patterns themselves. The idea is that the
averaging process reduces the intensity of the noise,
which is in first approximation uncorrelated with the
ECG signal, ensuring more precise measures. The
isoelectric level was identified following the standard EN
60601-2-51:2005-06 as having duration of more than 6
ms and amplitudes not exceeding 20 µV for at least three
samples. The Q onset was located at the end of the
isoelectric level preceding the QRS complex. The T offset
was instead detected using the Lepeschkin & Surawicz
method [14] as the point at the intersection of the
isoelectric level and the tangent of maximum slope to the
T complex. The tangents were built using a local least
squares approximation on short windows of 20 ms.
Finally, individual <strong>QT</strong> measures might be obtained for
each beat of the family adjusting the conventional Q start
and T end points, as used to build the average patterns,
with the offsets they show with respect to the patterns’ Q
<strong>QT</strong> (ms)
400
350
300
250
200
740 760 780 800 820
RR (ms)
Figure 2. <strong>QT</strong>/RR relationship for recording s0398lre.
Each point in the figure was obtained measuring the <strong>QT</strong>
interval for the first beat of each group of similar beats,
while the RR value was averaged within each group. The
dashed line is the least squared linear interpolant.
onset and T offset. For example, if the T offset was
detected on the average pattern to be 15 ms in front of the
T start points on which the beats were aligned, each of the
T start points might be moved of 15ms leading to an
estimate of each T offset.
3. Results
The method was employed to automatically measure
<strong>QT</strong> segments on the 549 diagnostic ECG recordings
which constitute the PTB Diagnostic ECG Database used
for the 2006 Computers in Cardiology Challenge. The
measures were performed on lead II, as required for the
challenge. Moreover, it was required to provide a <strong>QT</strong>
measurement for the first “representative beat” of the
recording. We considered as “representative” the first
beat of the first group of beats. In fact, if entering a group
at the end of the refinement process such beat was
guaranteed to be at least similar to a number of beat
following it.
Due to low signal quality (2 cases) or to a high heart
rate variability within the recording (which reduced the
possibility to have groups of at least 2 beats with |RR i+1 -
RR i |

situations in which the patterns arise from averaging only
a small number of beats, such as when, in short ECG
segments, the noise level is not negligible, the heart rate
variability high or many ectopic beats are present. In
those situations taking the measures on the average
patters is of little help unless we are able to increase the
number of beats in the group. Therefore, the use of two
separate patterns possibly increases the number of
recording on which effectively using the technique.
Unfortunately, the technique we suggested proved
capable of automatically refining the positions of Q
onsets and T offsets in only a fraction of the total
recordings (about 87%). When the recordings were highly
corrupted by noise or the initial guess for the end of the T
wave was misplaced, often the algorithm ended up
misleadingly providing an estimate of the end of the P
wave of the following beat (thus furnishing estimates for
the <strong>QT</strong> interval far outside the physiological range).
Obviously, for the sake of the competition, we could have
substituted all the <strong>QT</strong> we knew were outside the
physiological range (e.g. <strong>QT</strong>500 ms)
with a common physiologically-reasonable value, let’s
say 400 ms. As we verified after the competition was
ended and the “gold standard” scores published, doing so
our hypothetical score would have been of 44.63 ms,
much closer to what we consider the true performances of
our algorithm (see previous section). Nevertheless, we
considered that such substitution, meaningful only for the
competition, would have been of any scientific interest
and we rather avoided it.
The method we suggested needs surely further work
and a fine tuning. A multi-lead approach might be a
possible improvement: applying the method on more than
a lead could help to resolve those situations in which the
<strong>QT</strong> value computed on the lead II was outside the
physiological range. Also, a rough refinement of the
initial guess for “T end” might help.
No matter how the method we proposed scored, we
feel that the overall competition showed that it might be
possible in the future to accept automated methods among
those trusted to measure the <strong>QT</strong> interval from ECG
recording, thus also ensuring a higher repeatability of
measurements than today.
Acknowledgements
The authors thank the organizers of the challenge for
the effort shown in organizing the competition.
References
[1] de Bruyne MC, Hoes AW, Kors JA, Hofman A, van
Bemmel JH, Grobbee DE. Prolonged <strong>QT</strong> interval predicts
cardiac and all-cause mortality in the elderly - The
Rotterdam Study. European Heart Journal 1999;20:278–
284.
[2] Montanez A, Ruskin JN, Hebert PR, Lamas GA,
Hennekens CH. Prolonged <strong>QT</strong>c <strong>Interval</strong> and Risks of Total
and Cardiovascular Mortality and Sudden Death in the
General Population. Aech Intern Med 2004;164:943-948.
[3] Gittleman IW, Thorner MC, Griffith GC. The Q-T interval
of the electrocardiogram in acute myocarditis in adults,
with autopsy correlation. Am Heart J. 1951;41:78-90.
[4] Lepeschkin E, Surawicz B. The measurement of the Q-T
interval of the electrocardiogram. Circulation;1952;6:378-
388.
[5] Hoffman JIE, Lister G. The implications of a relationship
between prolonged <strong>QT</strong> interval and the Sudden Infant
Death Syndrome. Pediatrics 1999;103:815-817.
[6] Roden DM. Drug-Induced Prolongation of the <strong>QT</strong> <strong>Interval</strong>.
N Engl J Med 2004;350:1013-1022.
[7] Bazett HC. An analysis of the time-relations of
electrocardiograms. Heart 1920;7:353-370.
[8] Malik M, Färbom P, Batchvarov V, Hnatkova K, Camm
AJ. Relation between <strong>QT</strong> and RR intervals is highly
individual among healthy subjects: implications for heart
rate correction of the <strong>QT</strong> interval. Heart 2002;87:220-228.
[9] ICH. E14 Clinical Evaluation of <strong>QT</strong>/<strong>QT</strong>c <strong>Interval</strong> Prolongation
and Proarrhythmic Potential for Non-Antiarrhythmic
Drugs. 2005. http://www.fda.gov/cber/gdlns/iche14qtc.htm
[10] Murray A, McLaughlin NB, Bourke JP, Doig JC, Furniss
SS, Campbell RW. Errors in manual measurement of <strong>QT</strong>
intervals. Br. Heart J. 1994;71:386-390.
[11] Badilini F, Maison-Blanche P, Childers R, Coumel P. <strong>QT</strong>
interval analysis on ambulatory electrocardiogram
recordings: a selective beat averaging approach. Med Biol
Eng Comput. 1999;37:71-9.
[12] Moody GB, Koch H, Steinhoff U. The PhysioNet /
Computers in Cardiology Challenge 2006: <strong>QT</strong> <strong>Interval</strong>
Measurement. Computers in Cardiology (33) 2006.
[13] Laguna P, Jané R, Bogatell E, Anglada DV.
ECGPUWAVE, freely available from www.physionet.org.
[14] Lepeschkin E, Surawicz B. The measurement of the Q-T
interval of the electrocardiogram. Circulation 1952;6:378-
388.
Address for correspondence
Roberto Sassi
Dipartimento di Tecnologie dell'Informazione
Università degli Studi di Milano
via Bramante 65, 26013 Crema (CR) Italy
E-mail address: sassi@dti.unimi.it